[PATCH] Loop Vectorizer: Refactor code to compute vectorized memory instruction cost

[Renato Golin]

Hi Arnold,

I like the approach, and I think we could do the same for all types of
instructions, not just load.

Regarding this specific code, I have some comments...

1. You seem to be cramming load & stores with a lot of non-shared code.
Maybe it'd be more efficient to have a MemoryCostComp that both
LoadCostComp and StoreCostComp derive and implement the specific parts.

2. Assert Liberally!

+    // It better be a load now.
+    Load = cast<LoadInst>(MemInst);

to

+    Load = dyn_cast<LoadInst>(MemInst);
+    assert(Load && "Invalid memory instruction type");

or use llvm_unreachable().

3. All this can be omitted if you use Instr* instead of specific
LoadInst/StoreInst and if you common up code in a base class:

+  unsigned getOpcode() {
+    if (Store)
+      return Store->getOpcode();
+
+    return Load->getOpcode();
+  }
+
+  unsigned getAlignment() {
+    if (Store)
+      return Store->getAlignment();
+
+    return Load->getAlignment();
+  }
+
+  unsigned getPointerAddressSpace() {
+    if (Store)
+      return Store->getPointerAddressSpace();
+
+    return Load->getPointerAddressSpace();
+  }



cheers,
--renato
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20130204/664ade84/attachment.html>